Professor Daphne Koller

Computer Science Department, Stanford University
Email: koller@CS.Stanford.EDU
Web: click here

My main research focus is on dealing with complex domains that involve large amounts of uncertainty. My work builds on the framework of probability theory, decision theory, and game theory, but uses techniques from artificial intelligence and computer science to allow us to apply this framework to complex real-world problems.

Most of my work is based on the use of probabilistic graphical models such as Bayesian networks, influence diagrams, and Markov decision processes. Within that topic, my work touches on many areas: representation, inference, learning, and decision making. One main focus has been the extension of the representational power of the probabilistic graphical modeling language, to encompass a much richer set of domains. For example, work in my group includes:

• incorporating hierarchical and object-relational structure in our object-oriented

Bayesian networks (OOBNs) and probabilistic relational models (PRMs);

• extensions to temporal domains using dynamic Bayesian networks;
• hybrid Bayesian networks involving both discrete and continuous variables;
• factored MDPs that represent sequential decision problems in a factored way;
• structured representations for utility functions;
• multi-agent influence diagrams for representing multi-agent decision problems with incomplete information;

I believe that a good representation must also support effective inference and learning algorithms. Hence, the work done in my group is also highly focused on these topics. We have worked on exact and approximate inference algorithms for these representations, and on approaches for learning these models from data. On the inference side, we have done a lot of work on inference in dynamic Bayesian networks, inference in hybrid Bayesian networks, decision making in factored MDPs, and inference for large scale models such as those generated by a PRM or an OOBN. On the learning side, we have done a lot of work on learning probabilistic models from relational databases, on active learning of probabilistic models (where the learner can query for particular types of instances), and on learning utility functions from data. Our work spans the range from concepts to theory to applications. Some of our work is conceptual: defining new representation schemes and exploring their expressive power. Some of it is theoretical and algorithmic: designing new inference and learning algorithms and proving that they achieve certain properties. And some is applied: experimenting with our approaches on both synthetic and real problems. Some of the applications that we are particularly interested in right now are: learning models from rich heterogenous biomedical databases, which can include clinical, genomic, genetic, and epidemiological data; fault diagnosis for complex hybrid systems; and tracking at the symbolic level from low-level visual data.

Some of the main funding sources for my work include:

• From Sensors to Symbols: An Integrated Approach: a Presidential Early Career

Award for Scientists and Engineers (PECASE), via the Office of Naval Research Young Investigator Program.

• Probabilistic Relational Models for Biological Data: an NSF ITR award.
• Decision Making under Uncertainty: MURI project with U.C. Berkeley and U.C. Davis.
• An Integrated Approach to Intelligent Systems: MURI project with U.C. Berkeley and Cornell.